Independent case study | one locally run GPT Researcher workflow
What ANAGANI found in one research run
We analyzed an open-source GPT Researcher workflow that we ran ourselves. This independent review is not affiliated with or endorsed by its maintainers.
Explore the engineering findings
These are investigation leads from one trace, not confirmed defects. The report's 12,000-character threshold flags retrieval output size; it is not a recommended prompt limit.
Large retrieval outputs | RAG-002 | high change risk
The saved Engineering Report records four normalized retriever outputs ranging from 69,724 to 162,409 characters, above its configured 12,000-character detection threshold. Those output sizes are not the size of the final writer prompt.
Next decision: determine which retrieved documents reach the final prompt and contribute to the answer. Compare correctness and grounding before and after any smaller retrieval configuration. Output size alone does not prove wasted prompt tokens.
Input token composition | TOK-001 | medium change risk
The workflow used 21,415 measured input tokens; the report-writing LLM call accounted for 19,193 of them.
What we checked: a local, content-free inspection measured 74,573 characters in the writer's user message. Its size points to an evaluation opportunity; it does not show that its sources are irrelevant. Any shorter-context change needs answer-quality and latency checks.
Intermediate LLM orchestration | LLM-001 | medium change risk
The two earlier LLM calls served distinct purposes: one established the research role and the other generated search queries used by the workflow. This run does not show either call was unnecessary.
Next decision: change either step only if a replacement meets the application's quality requirements on representative research tasks.
What the trace cannot establish
Retriever spans recorded empty query fields, so matching input fingerprints do not establish repeated effective searches. Span gaps cannot be assigned to retrieval network time without better instrumentation. This run does not establish financial savings or answer quality.
Read the original ANAGANI reports
These reports show what ANAGANI produced for this run. We later checked the retrieval evidence: the export omitted search queries, so the report's repeated-search recommendation is unconfirmed. Check the actual searches before considering caching.
See what ANAGANI could investigate in your workflow
To discuss your workflow, request a guided pilot. We will agree on review scope and a trace you have sanitized before sharing.
Request a guided pilotScope: one execution. Recommendations need validation across representative runs before production changes. No optimization or savings are claimed.